Design and implementation of a reflective well-being app

Mood-tracking app scan support mental health self-management by helping users monitor their emotional states and identify patterns int heir daily lives. However, most existing apps focus on data collection and visualisation while offering little personalised guidance, and many raise serious privacy concerns by sharing sensitive data with third-party services.This thesis presents MoodScape, a reflective well-being app for Android that combines context-aware mood tracking with personalised, LLM-generated recommendations while preserving user privacy through a locally hosted open-source language model. MoodScape collects mood entries alongside contextual signals from weather, music, health, and social interaction APIs and uses a locally deployed Ollama model (llama3.2:1b) on a university-managed server to generate tailored recommendations—without transmitting any mood or health data to commercial cloud providers. The system was evaluated in a two-week AB/BA crossover field study with ten participants. Each participant experienced both a tailored recommendation condition, in which the LLM drew on their personal mood and context data, and a generic baseline condition, in counterbalanced order. Perceived recommendation quality, systemusability (SUS), user experience (UEQ), and self-reported reflection were assessed through standardised questionnaires. The results show a consistent descriptive pattern favouring the tailored condition: participants rated the LLM- generated recommendations higher than the baseline on all four comparison items, with the largest difference on perceived personalisation (𝑀 = 2.80 vs. 𝑀 = 2.00), which also reached statistical significance in a supplementary Wilcoxon signed-rank test (𝑝 = .039, 𝑟 = .65). However, absolute satisfaction levels remained moderate across both conditions, and engagement with the app varied considerably across participants (4–38 mood entries). MoodScape achieved above-average usability (𝑀SUS = 74.75) and positive user experience scores, with participants particularly valuing the data exploration features. Reflection items indicate that the app supported data exploration and, to a degree, self-reflection, though deeper behavioural change did not emerge within the study period. These findings demonstrate that privacy-preserving, locally hosted LLMs are technically viable for generating mood-related recommendations in a GDPR-compliant architecture and that the personalisation approach itself is perceived positively. The quality gap compared to larger commercial models, however, constrained the practical impact. As open-source language models continue to improve, the approach demonstrated by MoodScape offers a promising path towards AI-powered well-being support that respects user privacy.